Deep neural networks with promising diagnostic accuracy for the classification of atypical femoral fractures

Georg Zdolsek; Yupei Chen; Hans-Peter Bögl; Chunliang Wang; Mischa Woisetschläger; Jörg Schilcher

doi:10.1080/17453674.2021.1891512

Authors

Georg Zdolsek Department of Orthopedics and Department of Biomedical and Clinical Sciences, Faculty of Health Science, Linköping University, Linköping
Yupei Chen Department of Biomedical Engineering and Health Systems, Royal Institute of Technology, Stockholm
Hans-Peter Bögl Department of Orthopedics and Department of Biomedical and Clinical Sciences, Faculty of Health Science, Linköping University, Linköping; Department of Orthopedic Surgery, Gävle Hospital
Chunliang Wang Department of Orthopedics and Department of Biomedical and Clinical Sciences, Faculty of Health Science, Linköping University, Linköping
Mischa Woisetschläger Department of Radiology and Department of Medical and Health Sciences, Linköping; Center for Medical Image Science and Visualization, Linköping University, Linköping
Jörg Schilcher Department of Orthopedics and Department of Biomedical and Clinical Sciences, Faculty of Health Science, Linköping University, Linköping; Wallenberg Centre for Molecular Medicine, Linköping University, Linköping, Sweden

DOI:

https://doi.org/10.1080/17453674.2021.1891512

Abstract

Background and purpose — A correct diagnosis is essential for the appropriate treatment of patients with atypical femoral fractures (AFFs). The diagnostic accuracy of radiographs with standard radiology reports is very poor. We derived a diagnostic algorithm that uses deep neural networks to enable clinicians to discriminate AFFs from normal femur fractures (NFFs) on conventional radiographs.

Patients and methods — We entered 433 radiographs from 149 patients with complete AFF and 549 radiographs from 224 patients with NFF into a convolutional neural network (CNN) that acts as a core classifier in an automated pathway and a manual intervention pathway (manual improvement of image orientation). We tested several deep neural network structures (i.e., VGG19, InceptionV3, and ResNet) to identify the network with the highest diagnostic accuracy for distinguishing AFF from NFF. We applied a transfer learning technique and used 5-fold cross-validation and class activation mapping to evaluate the diagnostic accuracy.

Results — In the automated pathway, ResNet50 had the highest diagnostic accuracy, with a mean of 91% (SD 1.3), as compared with 83% (SD 1.6) for VGG19, and 89% (SD 2.5) for InceptionV3. The corresponding accuracy levels for the intervention pathway were 94% (SD 2.0), 92% (2.7), and 93% (3.7), respectively. With regards to sensitivity and specificity, ResNet outperformed the other networks with a mean AUC (area under the curve) value of 0.94 (SD 0.01) and surpassed the accuracy of clinical diagnostics.

Interpretation — Artificial intelligence systems show excellent diagnostic accuracies for the rare fracture type of AFF in an experimental setting.